Answers are grounded in this paper and cite their sources.
This thesis explores improved crop yield prediction in agriculture using domain-informed learning techniques, highlighting predictive reliability.
Driven by the increased availability of Earth Observation data, Deep Learning has been increasingly adopted in Earth Observation applications. At the same time, we observe a tendency towards larger Deep Learning models, powered by ever larger datasets. Such purely data-driven models are exceptionally powerful. However, scenarios exist where purely data-driven methods reach limits. For instance, when data is insufficient, when physical principles must be considered, or when information about the uncertainty of a prediction is required. All these aspects nourish skepticism and continue to hinder the success of Deep Learning approaches in Earth Observation applications. This thesis explores the inclusion of prior knowledge into a learning system, defined as Domain-Informed Learning. Here, prior knowledge is a source of information that exists independently of the model. This work focuses on agricultural applications and time series analysis, and develops methods that are extendable to a variety of Earth Observation tasks. This work introduces a novel, large-scale dataset for crop yield prediction using Earth Observation data. Following, three techniques of integrating prior knowledge are explored, namely 1) data space enrichment, 2) conditional learning, and 3) uncertainty estimation. The first technique analyzes data sources and time series representations for crop yield prediction. Furthermore, novel data fusion methods are presented. In the following, this work analyzes conditional learning with prior knowledge. A novel physics-guided approach for drought stress estimation is proposed. The last part emphasizes the inherent variability in Earth Observation tasks. Therefore, the concept of uncertainty estimation is introduced by focusing on missing data and distribution shifts. We present a novel method for uncertainty estimation, inspired by naturally occurring missing time steps. Finally, we overcome the performance collapse under distribution shift by coupling Bayesian inference with prior knowledge. In conclusion, this thesis contributes to the field of research by making models more reliable, easier to understand, and more trustworthy. It offers new perspectives on Earth Observation and emphasizes the importance of understanding how confident our predictions are and that they remain consistent with real-world physical laws.
No takes yet. Share an insight, caveat, or question.
Miro Benjamin Miranda Lorenz (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: